Slumlords of the Intelligence Age: How the AI Debt Boom Could Go Bust

The State of Drift Part 2: The machines that wear out, the product that melts, and the people left holding the bill.

Jonathan Gordon

•

Closeup of a brick wall with a large vertical crack

The State of Drift is a series about a quiet problem in software. Software drift is the accumulation of code mistakes that still appear to work (Gordon, 2026). It isn't new. AI didn't create drift on its own. It just speeds it up past the point where teams can keep up.

TL;DR

  • The AI boom is being built largely on borrowed money, secured by chips that wear out in a few years, to make a product whose price falls about 80% a year.

  • The technology is real, and so is the collapse in what it sells for. Neither of those has a due date. The debt does.

  • The crash doesn't need AI to fail. It only takes one year in which revenue grows more slowly than prices fall.

  • Two things turn the crash into a clearance sale instead of a disaster: machines we can trust, and a market that can see failures clearly. We have neither yet, and we should start building both now.

A note on some of the people in this piece: none of them are composites. They're real, they said these things on the record, and they're all in the sources at the end. I've left their names out on purpose. Names pull your attention toward personalities, and I care more about the seat each person sits in, because anyone in that seat would be staring at the same math.

Where this began

A friend sent me a podcast with Paul Kedrosky, an investor and writer who has watched money flood into big ideas before. He argues the AI boom will end the way building booms end: in a glut. Not because the technology fails, but because the money gets spent before the truth arrives. Capacity gets poured on the promise of use that hasn’t shown up yet.

I listened to it twice. That line, “spent before the truth arrives,” stuck with me, because I’d heard its quieter echo for most of my career.

I’ve spent thirty years making software, and I’ve watched the same shape play out at desk scale. I call it drift: the small mistakes and omissions that pile up inside code that still runs. It all looks like it’s working. The code ships on a promise, that it does what it claims, and only later, when real users lean on it, does anyone find out where it doesn’t.

Same move, two sizes. A feature gets written, a data center gets poured; each is a bet placed before the evidence is in, and the bill for the gap between the promise and the reality comes due later, when something finally has to bear weight. The boom isn’t drift. Drift lives in code. But it is the same shape, poured in concrete and capital: commitment before evidence, a surface that still looks fine, and a reckoning that waits until someone leans on it. The small version I’ve spent years naming. The big one is what Kedrosky calls a glut: a whole industry’s worth of things that looked fine until someone had to depend on them.

All that time, I’ve wondered how far a codebase can drift and still be worth something. Kedrosky got me asking the same question about the whole boom.

The warning at the center of this piece is his, and I’ll say so whenever I’m leaning on him. The way of seeing it, through drift, is mine.


The factory: Data centers

The man who sells the machines (the chips every AI system runs on) wants you to picture a factory. In his telling, a data center isn't a warehouse full of computers. It's a plant, and what it makes is "tokens": the words, images, and code AI produces, sold by the batch.

Morgan Stanley puts the buildout bill at about $3 trillion through 2028, and much of that money has gone into the factory picture. I understand why. It's a profitable picture. But it also turns the largest corporate construction spree in history into production capacity, and production capacity is only worth what its product sells for.

Keep that in mind. Everything else in this piece comes back to it.


The landlords: GPU cloud companies and the debt behind them

Alongside the tech giants, a new kind of landlord has shown up. These operators lease land and power, borrow against the chips themselves, and rent out the capacity. Kedrosky describes data centers as "really just multi-tenant apartment buildings. It just so happens there's no humans in the apartment building, there's just GPUs."

The catch is that apartment buildings last. Chips running around the clock can fail within 18 months, and the accounting assumes they'll last five or six years. The man who sells them once joked that when the next generation ships, "you couldn't give away" the last one. The loans, meanwhile, are written like mortgages, on a decades-long timeline.

Rent out a box you know will rot, book it as if it won't, and arrange things so that when it fails, someone else is holding the lease. That is a slumlord's business model, and it is the one this boom is built on. A slumlord’s whole edge is the gap between a building that still collects rent and a building that’s sound. Like drift in a codebase, the rot hides under a surface that still works, only here it’s poured in concrete and financed like a mortgage.

If a tenant walks out of an apartment building, you still have a building. If a tenant walks out of a data center, you're left with a warehouse of aging electronics and a power contract. And the landlords are the smaller problem. The bigger one is what their factories make.


The product melts: AI prices are falling about 80% a year

AI output keeps getting cheaper, about 80% a year by industry trackers' count, and nobody expects it to stop. A batch that cost $30 in early 2023 runs about 14 cents now.

The usual answer is that cheaper just means everyone uses far more of it, and so far that's held. The bills keep climbing. But follow the price down the way Kedrosky does. A dollar of output is worth 20 cents a year from now, 4 cents the year after, under a penny the year after that. Just to stay in place, you have to sell five times as much every year. To get ahead, more than that, and again the next year, and the next. It's a down escalator that keeps speeding up.

And you can't win by being the best, because nobody can tell whose AI is whose. Kedrosky's test is simple: show people two models and ask which is which. "Everyone thinks they can tell the difference. And the reality is no one can tell the difference."

That makes it a Pepsi Challenge, as he puts it. The label is the only real difference, and the price keeps dropping. You can sell more cans every year for less and get away with it for a good while, right up until someone asks you to pay off the loan on the bottling plant.


Other people's money: Who is financing the AI buildout

So whose money is riding on all this? Mostly not the builders'. The Bank for International Settlements found that by mid-2026, more than half the money going into data centers was "external financing", the polite term for loans, bonds, and private credit.

The chipmaker, for its part, has lined up something like half a trillion dollars with the world's biggest asset managers to bankroll the very customers who then buy its chips. Critics call that circular financing: the money leaves by one door and comes back as sales through another. And if it goes wrong, the losses are built to land on ordinary bond investors, debt that ends up tucked inside insurance companies, pension funds, and index funds, which is to say, inside the savings of people who've never once thought about a GPU.

None of this means the technology fails. A boom like this doesn't break because the thing stops working. It breaks when there's more of it for sale than there are customers to buy it. The most convincing timeline I've seen puts that glut around 2029, when enough new capacity finally comes online to show whether the demand is really there. For now, everything's sold out. But that's just what the middle of a building boom looks like, and it tells you almost nothing about how it ends.


Why are they selling it to us?

If this technology is about to be smarter than all of us, why are the companies that make it so eager to rent it out instead of keeping it? Kedrosky puts it more bluntly: "If it can eat the whole economy, why not just ingest the economy and stop selling it to us?"

I can think of three honest answers.

The lead won't keep. When prices fall 80% a year, intelligence is perishable inventory. Renting it out fast is how you sell it before it spoils.

The bills are due. The chipmaker invoices every quarter, and the technology can't pay for itself yet. It needs customers: their money, their data, and their trust.

The sale is the strategy. This is the darkest reading, and it comes from the CEO of a company that wrote the playbook on smothering its own ecosystem. He says business customers are "paying twice": once in cash, and again in the know-how their questions teach the machine.

You can already see that third one playing out. AI companies are moving into their customers' industries, design and finance and healthcare among them, and turning businesses that were built on top of them into features.

Whichever reading you believe, watch what they stop selling, and when. The day the best model is no longer for rent, kept at home for its maker's own use, is the day they finally believe their own pitch.


The machines don't run true: Why AI errors compound

These machines are very good guessers. Ask one the same question twice, and you may get two different answers. That's fine for a poem, but it's a real problem for the systems that run banks, hospitals, and payroll, which need the same answer every time.

And the guessing compounds. A step that's right 95% of the time, chained 20 times, gets the whole job right only about a third of the time. Knowing that, I'm not surprised that something close to 95% of corporate AI pilots never deliver. This is drift again, out in the open: small errors that each look harmless, stacked until the finished work can't be trusted. And this is where drift meets the money. The crash only needs one year when revenue lags the falling price, and a product nobody can trust is how revenue lags. Every failed pilot is drift showing up as money that never arrives, before the bond market ever says so out loud.

I keep coming back to the textile mills for this, because they tell the story better than any white paper. The building we've been calling an apartment block is, on the inside, a factory floor. The factory system started in spinning. The first thing the Industrial Revolution mechanized was thread, before cloth. In this picture, the data center is the mill, the chips are the spinning frames, and AI output is the thread.

Nobody wears thread, though. AI output only matters once it's woven into finished work that holds up: software that works, accurate reports, decisions someone can defend. The weaver, the skilled person who turns cheap thread into cloth worth wearing, is exactly the person being moved away from the loom.

Two centuries ago, the trade had a word for cheap cloth made without a craftsman's judgment: slop-work. We use a version of it today for AI-generated work that looks polished but doesn't actually move the task forward, "workslop." Researchers at Stanford and BetterUp found that each piece of it costs the person who receives it about two hours to clean up. Same story, faster clock.


The humans at the looms: What the boom is doing to workers

Follow the thread out of the mill, and it ends in someone's hands. That's where the cost shows up.

In September, an engineer's post went viral, and it came from a community of AI enthusiasts, of all places. He described whole teams working 12-hour days "just to press enter," approving machine output nobody reads, because management had decided that producing more was no longer the bottleneck.

Nobody feels proud when the work ships, because everyone assumes the machine did it. Some workers have started calling themselves meat proxies: people kept in the chain to absorb the blame when something breaks, since the machine can't be held accountable. Their judgment isn't what they're there for.

That same week, in another thread, an engineer in Scotland described watching his employer go under. The company had made server parts for 30 years, and it failed because AI data centers bought up the memory its customers needed. Others in those threads included a translator now paid a penny a word to fix machine translations, a visual-effects artist thinking about electrician school, and a scientist who can't buy computers because the boom already bought them.

Up close, this is what "AI taking over" looks like. The human is still there to answer for the results, even though their judgment is removed from the process.

There's an odd kind of hope in those threads too, from people with nothing to sell: somebody will have to be paid to clean up what the machines shipped. I think they're right. Fixing that mess is real, skilled, paid work. It may turn out to be the jobs program hiding inside the crash.


The wager: Superintelligence vs. the debt schedule

Whenever these numbers come up, there's one answer that's supposed to end the conversation: this is all the road to superintelligence, and you can't overpay for a miracle. Kedrosky's translation is that it's like being asked what you'd pay for a call option on immortality. "Mathematically, I should be willing to spend anything."

I'd watch what the believers do more than what they say. The executive selling the most compressed, glorious version of the future is the same one who warned that if his growth forecast is off by a single year, his company goes bankrupt. The biggest investors tell the miracle story to their own backers, then price these projects like real estate once the doors are closed.

A payoff in 2035 can't cover a debt due in 2029. Eternity has never made a loan payment on time.


The lenders move first: What the bond market is signaling

Kedrosky's warning on the podcast was about that debt. Most of the buildout isn't the builders' own money. It's borrowed, and those loans have piled up into "the largest piece of the high-yield bond market." If they go bad, the loss doesn't stay with the builders. It travels into insurers, pension funds, and the index fund in somebody's retirement account, until, in his words, "you're in it whether you like it or not."

He called it. Not long after, Bloomberg News reported that the bonds of one of the biggest GPU landlords fell to around 80 cents on the dollar, the price of debt people are afraid won't be paid back, while the company still reported a backlog in the hundreds of billions and its chief executive called the sell-off irrational.

That's the tell. The backlog is a promise. The bond is real money that can vanish. When the two disagree, I trust the one that can bleed.

And it wasn't one company's bad week. In the same stretch, the firms building AI with their own money barely moved, while the borrowers, the ones renting out chips on debt, fell as a group. The market was sorting the field by who owed money.

From Sept. 14 to Oct. 14, 2026, shares of six debt-funded GPU landlords fell 24% on average, while four self-funded Big Tech builders rose 4%. Source: Stooq.com (2026); author's calculations.

None of this says AI is going away. The technology is real, and so is the demand. But the clock is real too, and a signal this early isn't cause for panic. It's a cue to act, to build the two things that turn a crash into a clearance sale while we still hold the reins.


The morning after: How an AI crash could land softly

Booms have ended usefully before. The dot-com crash wiped out the investors, but the fiber-optic cable they paid for stayed in the ground. Other people bought it later for pennies, and it became the internet’s backbone. That’s what a soft landing looks like: the wreckage stays usable. Fiber could wait because it kept working, exactly as good years later as the day it was laid. Output that drifted can’t wait that way, and neither can judgment that’s been trained out of people. That’s the line between wreckage you can reuse and wreckage you can’t, and it’s why the fix is a machine that runs true.

This boom is set up against a soft landing in two ways. The chips wear out too fast to sit in bankruptcy court waiting for a buyer. And the losses are hidden, spread through private funds and repackaged debt where nobody can tell who's exposed. Kedrosky says it's "literally following the same playbook" as 2008, "just with a different asset class."

So the landing has to be built ahead of time, and it has two parts.

The first is trust machinery: tools that check the machines' work, give the same answer every time, and catch flaws before they ship. That's what makes nearly free AI output usable, and it puts a price back on human judgment instead of squeezing it out. The weavers had a phrase for a loom that kept every thread aligned. They said it ran true. That's what's missing right now, and it can be built. (It's also what I spend my days building, so weigh my optimism accordingly.)

The second is honesty about the money: debts out in the open, chips that can be resold, losses that can be priced. A loss with a price on it can be absorbed. A hidden one is how one industry's crash becomes everyone's.


Bottom line

Step back and look at the whole factory. The machines wear out in five years. The product loses 80% of its value every year. The work coming off the looms is full of flaws nobody checks. The owners are moving out of the building while they sell you the lease. And the promise of superintelligence keeps the whole set of books looking justified.

Most bubbles are built on a true story about the future, sold at the wrong price, to the wrong people, on the wrong schedule. Guessing when this one pops is a parlor game. The useful work is making sure two things exist when it does: machines that run true, and a market that can catch what falls.

The last crash left us cheap fiber and a decade of internet abundance. This one could leave us intelligence too cheap to meter. Or it could leave a trillion dollars of dead electronics and a generation trained not to use its own judgment. The future will probably get here. The bills will get here first, and the runway between them is ours to build.

—————————————————

JONATHAN GORDON is the Founder & CEO of ReWeaver AI. With nearly three decades of experience, he has shaped developer tools and enterprise software at Google, Apple, Microsoft, Oracle, and SAP. He holds two patents and specializes in human-centered design for complex systems, AI/ML integration, and developer tooling.


FAQs

What is software drift?

Software drift is the accumulation of code mistakes that still appear to work. It isn't new, but AI speeds it up, because code gets written and shipped faster than anyone can check it. The flaws usually surface only when real users depend on the software.

How could the AI bubble burst?

The most likely ending is a glut: more data center capacity than paying customers. AI doesn't have to fail for that to happen. Because AI prices fall about 80% a year and the chips wear out in a few years, it only takes one year in which revenue grows more slowly than prices fall, while the debt is still due.

Who is financing the AI data center buildout?

Mostly borrowed money. The Bank for International Settlements found that by mid-2026, more than half the money going into data centers came from external financing: loans, bonds, and private credit. If those loans go bad, the losses can spread to insurers, pension funds, and index funds held in ordinary retirement accounts.

Why does the lifespan of AI chips matter?

AI data center loans are written on long timelines, and the accounting assumes chips last five or six years. Chips running around the clock can fail much sooner, and each new generation makes the last one less valuable. Unlike fiber-optic cable after the dot-com crash, worn-out chips can't sit idle for years waiting for a buyer.

What would a soft landing for the AI boom look like?

Two things would turn a crash into a clearance sale. The first is trust machinery: tools that check AI's work, give consistent answers, and catch flaws before they ship. The second is honesty about the money: visible debts, resellable chips, and losses that can be priced before they spread.

Sources

From Maker to Minder: AI Coding and the Old Story of Lost Craft ›